The Research And Application Of Rock Debris Image Segmentation Based On Deep Learning

The rock debris logging technology is a crucial technology in geological exploration.The purpose of rock debris logging is to analyze the rock debris samples,and the rock debris image segmentation is the basis of the rock debris sample analysis.At present,a large number of image segmentation algorithms have been proposed at home and abroad.The traditional image segmentation algorithms mainly segment the rock debris images based on features such as color,but these algorithms have obvious problems of poor universality and low accuracy.With the widespread application of deep learning,the segmentation algorithm based on convolutional neural network has improved the accuracy of image segmentation.In view of the problems existing in the above research,this thesis studies and improves the rock debris image segmentation algorithm based on deep learning.The main work includes the following aspects:1.The traditional image segmentation algorithms are analyzed at home and abroad,and three traditional segmentation algorithms are adopted to process the rock debris image and analyze the results.2.In view of the low accuracy of rock debris image segmentation in traditional segmentation algorithms,the Attention U-net network is studied,and the improved FCA＿U-net is used for rock debris image segmentation.Firstly,the residual network Res Net101 is used as the encoder of Attention U-net network,which reduces the loss of feature information in the encoder process;Then the feature connection stage is improved by adding full-scale connection and atrous spatial pyramid pooling,so that the decoder stage could make full use of the multi-scale feature information,and the attention module was introduced to enhance the weight of the target region;finally,the fully connected conditional random fields are used to carry out post-processing operation on the rock debris image,and the relationship between pixels is used to refine or correct the rough or wrong segmentation areas.In view of the problem of insufficient lighting at the edge of the rock debris due to overlapping or close proximity of rock debris in the rock debris image,an image enhancement algorithm is added to improve the inaccuracy of labeling in the process of data set production.Experiments show that the improved FCA＿U-net network has a greater improvement in the accuracy of rock debris image segmentation than the Attention U-net network.3.In view of the problem that the rock debris image segmentation algorithm based on FCA＿U-net network cannot segment overlapping or close rock debris of the same type in rock debris images,this thesis proposes a rock debris image segmentation algorithm based on the improved Mask R-CNN network.The feature extraction network of the Mask R-CNN network is improved by adding a branch from low-level features to high-level features,so that the extracted features can fully combine the content description information and semantic information of multi-layer features;a new convolutional block feature pyramid networks branch is added,in which the feature fusion method is connected according to the channel direction,and the attention module is added in the fusion process,so that the segmentation and branch results of the target region are more accurate.In view of the problem of undersegmentation of rock debris edges in the segmentation results,post-processing based on morphological expansion is introduced to improve the accuracy of segmentation.Finally,the experimental results show that the improved Mask R-CNN network has a certain improvement in the segmentation effect of the rock debris image,and the network can solve the problem that the same debris cannot be segmented when the same debris overlaps or is close together.4.Design and implementation of rock debris image segmentation and recognition system.Firstly,the demands of the rock debris image segmentation and recognition system are studied,and the framework of the system functional modules is designed according to the demands.Then the modules in the system framework were implemented in turn,the specific function module mainly includes file module that can collect the rock debris image through the acquisition equipment,segmentation module containing two methods of FCA＿U-net network and improved Mask R-CNN network,recognition module for recognizing the result of segmentation,postprocessing module for further processing of results,and export modules that can export segmentation and recognition results.Finally,the whole process from collecting rock debris to analyzing rock debris images is tested,and the test results show that the rock debris image segmentation and recognition system can realize the function of rock debris image segmentation and recognition,and can meet the purpose of the design system.